Developing software for lesion detection, false positive and false negative evidence' reduction in breast magnetic resonance imaging

The aim of the thesis is to develop a software-based breast lesion detection and classification system by using images taken from MRI system that is a commonly preferred system for breast cancer diagnosis. The developed system can be referred as to a software-based decision-support system for the specialists. Five main steps are performed to reach the given target, each of these steps includes several signal processing and image processing methods. Five steps performed in the presented thesis are database construction, breast lesion detection, lesion feature extraction, selection of the most effective features and decision steps, respectively. In database construction step, the most appropriate images taken from the MRI device are selected together with the specialist. In addition, a filtering-based preprocessing step is applied to the images to eliminate the possible artifacts. Then, a two-stage segmentation process is applied for breast lesion detection. The first stage is to detect breast region that may include lesion, and the second step is to obtain the lesion region from the breast region. Local adaptive thresholding connected component analysis, integral of horizontal projection and masking techniques are used for breast region detection. Individual and hybrid segmentation algorithms are applied to the images for lesion detection. In the thesis, 25 different metrics were used to analyze the success of segmentation process. In lesion feature extraction step, histogram, shape and texture features are calculated. Totally 92 features are determined for each lesion and the least effective features are discharged from the feature vector by using Fisher score method. The last step of the thesis is classification/decision step. In this step, K-nearest neighbor, support vector machines, random forest, naïve Bayes techniques are utilized. According to the achieved results, the developed software-based system provides 91±0,06% accuracy for lesion detection, 90,36±0,069% accuracy for separation of benign and malignant lesions.